Can My Firm Use Claude or ChatGPT on Client Files?

Firms ask this in several forms: whether the Claude Team contract covers them, whether uploading client files to ChatGPT is a breach, or whether they need to buy a server and run a model in the office. In each case the underlying question is whether the firm can defend its use of AI tools that can now read client files, which contain confidential and often sensitive personal data.

These tools can be used on client files lawfully. Whether a particular firm's use is lawful depends on the account tier the firm uses, a small number of settings, and the documents the law expects to be in place before processing of this kind begins.

The account tier

The consumer and business versions of these products are treated very differently in data protection terms, and much of the concern about AI and client data comes from confusing them.

Consumer accounts, meaning the free and personal plans of ChatGPT and Claude, may use what the user enters to train the vendor's models, subject to settings the user must find and change. There is no contract with the firm, and the firm cannot see what staff enter.

The business tiers are Claude Team and Enterprise, ChatGPT Team and Enterprise, and both vendors' API platforms. They do not use customer content for training by default, and both vendors offer processor terms in a Data Processing Addendum, which is the contract the UK GDPR requires between a firm and its processor. On paper, the leading vendors' business terms are strong, and the specific terms for OpenAI's API are set out in the setup guide.

The highest-risk arrangement in common use is therefore a member of staff using a personal account for client work, with no contract, consumer defaults and no oversight. The first step for any firm is therefore to establish which accounts its staff actually use.

Processing client files can be lawful

When a tool on a business tier reads a client file, the vendor processes that data as the firm's processor, under contract and on the firm's instructions, which is a recognised legal arrangement, the same as cloud storage or a practice management system. The law does not require firms to keep these tools away from client data.

The firm must, however, be able to show its basis for the processing: each purpose the AI serves needs a lawful basis and, where files contain health information, mental capacity material or other special category data, a condition under Article 9. These relate to the advice the firm is engaged to give, and the analysis must be carried out and recorded.

Settings can take data outside the contract

A sound contract is only part of the position, because the products include settings that can move data outside it. Feedback and rating features can send a whole conversation to the vendor, and on one leading product that conversation may then be stored for up to five years and used for training. Retention also varies by tier: on at least one leading business plan, saved conversations have no automatic expiry, and configurable retention is available only at enterprise level. Sharing and connector features can widen access further.

These settings are visible and can be changed quickly, but the defaults do not always match the contract, and the contract's protections apply in practice only if the configuration matches them. Whether it does can be checked in the firm's workspace, and the result should be recorded.

The documents the law expects first

Where using AI on client files applies a new technology to sensitive personal data, the law expects a data protection impact assessment to be completed before the processing begins, including the lawful basis analysis. Depending on the condition relied on, a short appropriate policy document may also be required. The privacy notice should describe the AI processing, the records of processing should include it, any transfer to the US under the Standard Contractual Clauses should be supported by the firm's own transfer risk assessment, and the rules for staff should be written down. Good practice is for staff to sign that instruction and for the engagement letter to explain what clients are told.

Firms often adopt the tools before the documentation is in place. Where that has happened, the appropriate response is to complete the documents as a priority and to close the obvious exposures immediately, after which the assessment will confirm whether the tools can remain in use as configured.

FCA-regulated firms

The FCA has said that it will supervise AI through its existing framework and has not introduced a separate AI rulebook. There is therefore no FCA rule against using these tools because they are AI. Its existing expectations on records, supervision, outsourcing and customer outcomes still apply to how the tools are used. Whether a firm's permissions and arrangements meet those expectations is a judgement for the person responsible for compliance, made on documented facts.

Is a local model needed instead?

In most cases a local model is not needed, because the steps above involve settings and documentation. They cost a fraction of what it takes to buy, secure and maintain a model server, and a local server would also usually mean running a less capable model than the ones staff already use. A local model is justified only in narrow cases, such as a client or regulator requiring that data never leaves the UK, or a vendor materially changing the terms on which the cloud arrangement relies.

Summary

The law does not prevent a firm from using Claude or ChatGPT on client files, and the vendors' business terms are sound. Where there is exposure, it lies in the difference between what the firm is doing and what its configuration and documents record, and both can be checked and then brought into line.